NVIDIA maintains a dominant position in the AI GPU market, holding 86% share in 2025, largely due to the widespread adoption of its Blackwell GPUs and the robust CUDA software platform. However, the AI chip ecosystem is diversifying as companies like OpenAI, in partnership with Broadcom, invest in custom ASICs designed for optimized inference workloads. These custom chips, initially for in-house use, aim to lower costs and reduce reliance on NVIDIA hardware, following similar strategies by Google, Amazon, and Meta with their own AI silicon.
Broadcom has emerged as a credible competitor, achieving $4.4 billion in AI revenue in Q2 2025, driven by its ASICs and networking products. Despite NVIDIA’s mature ecosystem, the rising use of custom chips signals a shift, especially as hyperscalers pursue cost-performance over peak specs. Amazon, Google, Microsoft, and Meta continue to invest in in-house silicon, adopting hybrid setups that balance NVIDIA GPUs with internal solutions for better efficiency and control.
ASIC adoption is expected to grow as chips become more energy efficient, secure, and tailored for specific AI workloads. Future improvements include enhanced security, AI integration, workload specialization, and benefits from technologies like CXL for memory pooling. Hyperscalers, foundries such as TSMC and Samsung, semiconductor firms like Broadcom and Marvell, and companies providing supporting infrastructure—including EDA software, memory, and cooling—stand to benefit from expanding ASIC demand.
NVIDIA’s announcement of the Rubin and Rubin Ultra GPUs with higher power densities is accelerating the adoption of advanced data center cooling. Companies such as CoolIT Systems, Asetek, Vertiv, and Schneider Electric are positioned to benefit from the shift to liquid and immersion cooling as hyperscalers upgrade infrastructure to support next-generation AI workloads.
